new_TAID / README.md
sttkw's picture
Dataset card (work in progress)
7af6c02 verified
|
Raw History Blame Contribute Delete
2.82 kB
metadata
pretty_name: new_TAID
license: cc-by-4.0
tags:
  - intrinsic-image-decomposition
  - terrain
  - atmosphere
  - synthetic

new_TAID

Work in progress — shards are uploaded while rendering runs (target: 16,000 samples = 4,000 terrains × 4 atmosphere/sun patterns).

Synthetic terrain intrinsic decomposition dataset rendered with Blender 4.5 (Cycles). Within one seed the terrain, textures, vegetation, water, season, clouds and camera are fixed; the 4 patterns differ in atmosphere (Rayleigh / Mie / ozone densities) and sun position.

Image formation

I = T ⊙ (DA ⊙ DS + SA ⊙ SS) + V
column meaning format
I image (clipped to [0, 1]) 16-bit RGB PNG, linear
Mask R = water, G = terrain, B = sky 8-bit one-hot RGB PNG
T camera → surface transmittance of the atmosphere (analytic from depth; clouds not included; 0 in sky) 16-bit RGB PNG, linear
DA diffuse albedo (not attenuated) 16-bit RGB PNG, linear
DS diffuse shading (HDR) NumPy .npy bytes, float16 HWC
SA specular albedo (not attenuated) 16-bit RGB PNG, linear
SS specular shading (HDR, OIDN-denoised) NumPy .npy bytes, float16 HWC
V in-scattered light of the atmosphere (clipped at 1 in bright sky) 16-bit RGB PNG, linear
Depth inverse depth normalised per image (planar Z): v = 0 sky, else z = 1 / (1/depth_max + (v-1)/65534 · (1/depth_min − 1/depth_max)) 16-bit gray PNG
depth_min, depth_max metric depth range of the image (m) float
seed, pattern, scene_id terrain id, atmosphere/sun pattern, "{seed}_{pattern}"
config all scene parameters, camera matrix, encodings text

In the sky (Mask B): T = DA = SA = 0 and V = I. Within a seed, DA and SA are identical across the 4 patterns.

Loading

PIL decodes 16-bit RGB PNGs to 8 bit, so keep the raw bytes and decode with OpenCV:

import io, cv2, numpy as np
from datasets import load_dataset, Image

ds = load_dataset("sttkw/new_TAID", split="train")
for c in ["I", "Mask", "T", "DA", "SA", "V", "Depth"]:
    ds = ds.cast_column(c, Image(decode=False))

def png(cell):
    a = cv2.imdecode(np.frombuffer(cell["bytes"], np.uint8), cv2.IMREAD_UNCHANGED)
    a = a[..., ::-1] if a.ndim == 3 else a              # BGR -> RGB
    return a.astype(np.float32) / (65535.0 if a.dtype == np.uint16 else 255.0)

row = ds[0]
I, T, DA, SA, V = (png(row[c]) for c in ("I", "T", "DA", "SA", "V"))
DS = np.load(io.BytesIO(row["DS"])).astype(np.float32)
SS = np.load(io.BytesIO(row["SS"])).astype(np.float32)
recon = np.clip(T * (DA * DS + SA * SS) + V, 0, 1)      # == I up to quantisation

Split train/validation by seed (the 4 patterns of a seed share the terrain and camera).

License

CC BY 4.0